Agent skill · research science · aiskillstore
logprob-prefill-analysis
Reproduces the full prefill sensitivity analysis pipeline for reward hacking indicators. Use when evaluating how susceptible model checkpoints are to exploit-eliciting prefills, computing token-based trajectories, or comparing logprob vs token-count as predictors of exploitability.
Why this skill is useful
Adds a complete executable pipeline for prefill sensitivity analysis, including scripts for checkpoint evaluation and data processing that are not commonly available.
What it needs
Requires djinn, vllm installed locally. About 5k tokens when loaded. Last updated 2026-08-07. 405 stars on the source repository.
What this skill does
Prefill Sensitivity Analysis Pipeline This skill documents the complete pipeline for measuring model susceptibility to reward hacking via prefill sensitivity analysis, including both token-based and logprob-based metrics. Quick Start: Single Command Reproducibility The full analysis can be run with a single command: This orchestration script: 1. Discovers checkpoints and prefill levels from the sensitivity experiment's config.yaml 2. Runs token-based trajectory analysis 3. Computes prefill logprobs for each checkpoint 4. Produces integrated analysis comparing token vs logprob metrics Overview The analysis measures how easily a model can be "kicked" into generating exploit code by prefilling its chain-of-thought with exploit-oriented reasoning. We track: 1. Token-based metric: Minimum prefill tokens needed to elicit an exploit 2. Logprob-based metric: How "natural" the exploit reasoning appears to the model Prerequisites Model checkpoints from SFT training Prefill source data (successful exploit reasoning traces) vLLM for serving checkpoints djinn package for problem verification --- Checkpoint Discovery The pipeline automatically discovers available checkpoints from a sensitivity experiment's config.yaml: The orchestration script reads this config to determine: Which checkpoints are available Which prefill levels were tested Where the SFT run directory is located --- Stage 1: Run Prefill Sensitivity Evaluation Evaluate each checkpoint at multiple prefill levels (0, 10, 30, 100 tokens). …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills aiskillstore/logprob-prefill-analysis